ChatGPT search cites pages that OpenAI's search systems can discover, fetch and judge relevant to the question being asked. If your site blocks OAI-SearchBot, OpenAI says it won't be shown in ChatGPT search answers, apart from plain navigational links. Being crawlable is therefore the first condition. After that, relevance, clarity and corroboration decide whether your page is the one quoted.
OpenAI doesn't publish a ranking formula, so anyone who claims to know the exact weights is guessing. What OpenAI does document is useful enough to act on.
The crawlers that matter
OpenAI lists its user agents on its bots documentation page. Three matter for most sites:
- OAI-SearchBot finds and surfaces websites for ChatGPT's search features. Block it and you opt out of ChatGPT search results.
- GPTBot collects content that may be used to train OpenAI's foundation models. Blocking it is a training opt-out and doesn't remove you from search.
- ChatGPT-User fetches a page when a user's request needs it, for example when someone pastes your URL. OpenAI notes that because these visits are user-initiated, robots.txt rules may not apply.
The settings are independent. A common choice is to allow OAI-SearchBot and decide separately about GPTBot. OpenAI says changes to robots.txt take about 24 hours to be reflected for search.
Check your CDN or firewall as well. Since July 2025 Cloudflare has asked new domains whether to block AI crawlers by default, and many site owners clicked yes without realising it covers search bots too. OpenAI publishes IP ranges for each bot, so you can allow them precisely.
How a source gets chosen
When a question needs fresh or specific information, ChatGPT rewrites it into one or more search queries, retrieves candidate pages and writes an answer that cites some of them. From the outside, a few patterns are consistent with that process:
- The page has to match a sub-question. A long question often becomes several searches. A page that answers one part precisely can be cited even if it doesn't cover the whole topic.
- Clear, quotable passages win. A paragraph that states a fact plainly, with a number or a named source, is easier to use than one that circles the point.
- Corroboration matters. If three independent sites agree on a fact, the answer is more likely to state it and cite one of them.
- The page must load without tricks. Content hidden behind heavy client-side rendering, cookie walls or login prompts may not be read at all.
The GEO research paper supports point two. In its tests, adding quotations, statistics and cited sources made a page's content appear more in generated answers, while keyword stuffing reduced it.
What you can change this month
- Allow OAI-SearchBot in robots.txt and at your CDN. Confirm it in your server logs (our post on reading server logs for AI crawlers shows how).
- Answer the question in the first two sentences of each key page, then expand.
- Add specifics. Replace "competitive pricing" with an actual price range. Replace "fast turnaround" with the number of days.
- Cite primary sources where you make factual claims. It helps readers, and it matches what the GEO research found.
- Be findable in Bing. ChatGPT search has historically drawn on third-party search partners as well as its own crawling, and Bing Webmaster Tools costs nothing to set up. Submitting your sitemap there is cheap insurance.
How to tell if it's working
OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral links from search results. In GA4, filter sessions by that source, or by the referrer chatgpt.com, and look at which landing pages receive the traffic. Expect the numbers to be small at first and messy at times; there have been short periods where referral data went missing.
For visibility without clicks, keep a fixed list of 20 to 50 prompts your buyers would use, run them monthly and record whether your site is cited. Our GEO service runs this tracking for clients, but a spreadsheet and an hour a month will get you started.
What not to do
Don't publish pages written only to be picked up by chatbots, and don't seed fake mentions on forums. Both tend to be caught by the same quality and spam systems that already police search. The pages ChatGPT cites most reliably are the ones a careful human editor would also pick: specific, sourced and easy to read.



